Bayesian mixture analysis for metagenomic community profiling.

Sofia Morfopoulou; Vincent Plagnol

The sensitivity and general applicability of metaMix comes at an increased computational cost, requiring access to a multi-core computer to run efficiently. For the datasets presented here, the computation time remained manageable and did not exceed a few hours, using 12 cores to run 12 parallel chains. Nevertheless, a limitation of metaMix is the increased processing time for very large datasets. Speed related improvements can be implemented in scenarios where the species ambiguity concerns only a small proportion of the read set. Reads with certain assignments can be flagged prior to the MCMC exploration of the state-space. Their assignment information can then be carried forward, thereby reducing the size of the similarity matrix used as input by the mixture model. Another area of possible improvement is MCMC convergence determination. The current version of metaMix produces log-likelihood traceplots allowing the user to visually inspect the MCMC convergence, however additional diagnostic criteria can be implemented in future versions. metaMix is most useful for complex datasets for which the interpretation is challenging. It has been mainly used as a clinical diagnostic tool, helping with the identification of the infecting pathogen while providing an accurate profile of the community in the sample.